app.py

import gradio as gr
import spaces
import os, torch, io
import json
os.system('python -m unidic download')
# print("Make sure you've downloaded unidic (python -m unidic download) for this WebUI to work.")
from melo.api import TTS
import tempfile
import wave
from pydub import AudioSegment


@spaces.GPU
def synthesize(conversation_text, speed, progress=gr.Progress()):
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    models = {
        'EN': TTS(language='EN', device=device),
    }
    speakers = ['EN-US', 'EN-Default']

    combined_audio = AudioSegment.empty()
    conversation = json.loads(conversation_text)
    for i, turn in enumerate(conversation["conversation"]):
        bio = io.BytesIO()
        text = turn["text"]
        speaker = speakers[i % 2]
        speaker_id = models['EN'].hps.data.spk2id[speaker]
        models['EN'].tts_to_file(text, speaker_id, bio, speed=speed, pbar=progress.tqdm, format='wav')
        bio.seek(0)
        audio_segment = AudioSegment.from_file(bio, format="wav")
        combined_audio += audio_segment

    final_audio_path = 'final.mp3'
    combined_audio.export(final_audio_path, format='mp3')
    return final_audio_path


with gr.Blocks() as demo:
    gr.Markdown('# Turn Any Article into Podcast')
    gr.Markdown('## Easily convert articles from URLs into listenable audio Podcast.')
    with gr.Group():
        text = gr.Textbox(label="Article Link")
    btn = gr.Button('Podcasitfy', variant='primary')
    aud = gr.Audio(interactive=False)
    btn.click(synthesize, inputs=[text], outputs=[aud])

demo.queue(api_open=True, default_concurrency_limit=10).launch(show_api=True)